Sign in Get started

Train a vision model for your exact task

Pick a task and a GPU. Pictograph converts your labels, trains on a managed GPU, scores the model, and hands back the weights. No infrastructure to manage.

Start free Read the guide 
# Freeze the dataset into an export, then train on a managed GPU
pictograph exports create road-signs --name v1 --format coco

pictograph train start road-signs v1 \
  --pipeline yolox \
  --gpu a10g \
  --name signs-detector

pictograph models download signs-detector --output signs.onnx
from pictograph import Client, TrainingRun

client = Client()

# Freeze the dataset into an export, then train on a managed GPU
client.exports.create(
    dataset_name="road-signs",
    name="v1",
    format="coco",
)
run: TrainingRun = client.training.create(
    dataset_name="road-signs",
    export_name="v1",
    pipeline_type="yolox",
    name="signs-detector",
    config={"epochs": 30},
)  # returns when the run finishes

client.models.download(
    name="signs-detector",
    output_path="signs.onnx",
)

From dataset to trained model

Pictograph converts your dataset to the right format, runs the job, and returns a model you can measure.

  1. Build a dataset

    Label in the editor, draft with SAM3 and review, or import from V7, Roboflow, COCO, YOLO, or VOC.

  2. Freeze an export

    An export records exactly which images and labels a run trained on, so any run can be repeated.

  3. Pick a task and a GPU

    Pictograph converts the labels for the architecture and trains on a managed GPU with live progress.

  4. Evaluate, then ship

    Score the model on held-out data, then deploy it, run it in a workflow, or download the weights.

Five vision tasks, and weights you own

Choose the architecture that fits the job. Every run produces a model you can evaluate, deploy, or run yourself.

Object detection YOLOX (nano to x) for speed at the edge, RF-DETR for accuracy.
Instance segmentation RF-DETR segmentation: a mask for every object.
Semantic segmentation Segformer, U-Net, or U-Net++ on ResNet or EfficientNet encoders.
Keypoints and pose RF-DETR keypoints, grouped per object for multi-instance pose.
Classification ResNet, EfficientNet, MobileNet, ConvNeXt, or ViT backbones.
Your weights Download ONNX or PyTorch weights, or build TensorRT and ExecuTorch versions for the edge.

Managed GPUs, priced in USD. Train on an A10G, A100, or H100, or let Pictograph pick the cheapest GPU the run fits on. Runs bill compute credits by the GPU-second, and you see an estimate before you start.

Keep exploring

Auto-annotation Draft labels with SAM3 point, box, and text prompts, and review them before you train. Deployments Serve your trained model at one authenticated /predict URL. Workflows Run your model over images, video, or a dataset in a visual pipeline.

Model training FAQ

What models can I train on Pictograph?

Object detection (YOLOX and RF-DETR), instance segmentation (RF-DETR), semantic segmentation (Segformer, U-Net, and U-Net++), keypoint and pose models (RF-DETR), and image classifiers (ResNet, EfficientNet, MobileNet, ConvNeXt, and ViT). Every model downloads as ONNX, and most also as PyTorch weights.

How much data do I need to train a model?

There is no fixed minimum, but more labeled images give a better model. With three or more images Pictograph holds some back for validation; with fewer it trains on all of them. Draft labels with SAM3 and review them, then let your own model pre-label the next batch.

How is training priced?

Training bills your compute credits by the GPU-second the run actually uses, on an A10G, A100, or H100. You see an estimate before you start and pay nothing for idle time. Training runs on every plan, Free included.

What happens after training finishes?

The model lands in your model registry with its validation metrics. From there you can evaluate it on another export, deploy it as an endpoint, run it in a workflow or a batch auto-annotation job, download its weights, or publish it on Explore.

Train your first model

$5 of free compute every month. No card required.

Start free Read the docs